Fig. 3: Benchmarking MODES over unimodal representations and other representation fusion strategies.

a Kernel regression on the proposed fused representations outperforms kernel regression on unimodal representations and other fusion strategies for predicting general physiological phenotypes (n = 4143, mean values are reported with error bars indicating one standard deviation). b Kernel regression on the proposed fused representations outperforms kernel regression on unimodal representations and other fusion strategies on various diagnostic tasks (n = 4143, mean values are reported with error bars indicating one standard deviation). c Kernel regression on the proposed fused representations outperforms kernel regression on unimodal representations and representations learned with the DropFuse model³ on predicting cMRI-derived phenotypes from ECG only, and predicting ECG derived phenotypes from cMRIs only. d Generated samples of one modality using reference samples of another modality. The first row shows the most probable ECG measures for a given cMRI sample and the second row shows what the cMRI would most likely look like given a reference ECG sample.